Circle Loss: A Unified Perspective of Pair Similarity Optimization

By Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, Yichen Wei. 2020

1,036 citations recorded by the source corpus.

This paper provides a pair similarity optimization viewpoint on deep feature learning, aiming to maximize the within-class similarity $s_p$ and minimize the between-class similarity $s_n$. We find a majority of loss functions, including the triplet loss and the softmax cross-entropy loss, embed $s_n$ and $s_p$ into similarity pairs and seek to reduce $(s_n-s_p)$. Such an optimization manner is inflexible, because the penalty strength on every single similarity score is restricted to be equal. Our intuition is that if a similarity score deviates far from the optimum, it should be emphasized. To this end, we simply re-weight each similarity to highlight the less-optimized similarity scores. It results in a Circle loss, which is named due to its circular decision boundary. The Circle loss has a unified formula for two elemental deep feature learning paradigms, \emph {i.e.}, learning with class-level labels and pair-wise labels. Analytically, we show that the Circle loss offers a more flexible optimization approach towards a more definite convergence target, compared with the loss functions optimizing $(s_n-s_p)$. Experimentally, we demonstrate the superiority of the Circle loss on a variety of deep feature learning tasks. On face recognition, person re-identification, as well as several fine-grained image retrieval datasets, the achieved performance is on par with the state of the art.

Subfields: Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

In Optimization & Training

Citations in the corpus

Margin-Based Few-Shot Class-Incremental Learning with Class-Level Overfitting Mitigation (2022, methodology): “As empirically proved by current works [33, 38, 11, 24], applying the margin-based classification could effectively improve the embedding space learned by metric-based methods.”
SphereFace Revived: Unifying Hyperspherical Face Recognition (2021, background): “The development of deep FR is also closely related to deep metric learning [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28].”
Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies (2021, background): “For example, in computer vision research, the feature representations, particularly the deep feature representation (He et al., 2020; Sun et al., 2020), are robust for machine perception.”
Geometric Transformer for Fast and Robust Point Cloud Registration (2022, methodology): “A straightforward so-lution is to adopt a circle loss [30] similar to [4,16].”
Channel Augmented Joint Learning for Visible-Infrared Recognition (2021, methodology): “A widely-used technique to measure the pair distance difference is ℓ 1 norm [29, 31, 34], as done Eq.”
Group-aware Label Transfer for Domain Adaptive Person Re-identification (2021, methodology): “In addition, we design a target instance memory bank combined with weighted contrastive loss [36] to make the model more powerful for features representation.”
A Metric Learning Reality Check (2020, methodology): “The tuplet margin loss [69] combines LogSumExp with an implicit pair weighting method, while the circle loss [52] weights each pair’s similarity by its deviation from a pre-determined optimal similarity value.”
ElasticFace: Elastic Margin Loss for Deep Face Recognition (2021, background): “35 (2) CircleLoss [26] (CVPR2020) clean MS1M [7, 26] 93.”
PREDATOR: Registration of 3D Point Clouds with Low Overlap (2020, methodology): “Circle loss: To supervise the point-wise feature descriptors we follow3 [3] and use the circle loss [34], a variant of the more common triplet loss.”
Self-Alignment Pretraining for Biomedical Entity Representations (2020, background): “Multi-Similarity loss (Wang et al., 2019) and Circle loss (Sun et al., 2020) are two recently proposed metric learning objectives and have been considered as SOTA on large-scale visual recognition benchmarks.”
Crossover Learning for Fast Online Video Instance Segmentation (2021, methodology): “…frame, respectively. p i ( n ) is optimized by cross-entropy loss: However, this approach suffers from the following issues (see Fig 3, left) [54, 47]: the feature space where e i and { e N } := { e 1 , e 2 , . . . , e N } live in is deﬁned by the sampled frames, and the decision boundary is…”
TransReID: Transformer-based Object Re-Identification (2021, methodology): “CNN-based methods [16, 23, 31, 37] have achieved great success in both person and vehicle ReID lately.”
Quasi-Dense Similarity Learning for Multiple Object Tracking (2020, background): “supervise dense matched quasi-dense samples with multiple positive contrastive learning by the inspiration of [42].”
FastReID: A Pytorch Toolbox for General Instance Re-identification (2020, methodology): “Therefore, we try to obtain the similar scores for all spatial features of X with respect to Y , and the final matching score can be defined as:”
WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (2021, background): “Even though growing efforts have been devoted to investigating sophisticated networks [9, 80, 8, 71, 16] and losses [32, 68, 66, 12, 27, 55, 10], academia is restricted by limited training set and nearly saturated test protocols.”
MISIM: An End-to-End Neural Code Similarity System (2020, methodology): “To train the models, we use the Circle loss [40], a state-of-the-art loss function that has been effective in various similarity learning tasks.”